block/elevator.c

   1 /*
   2  *  Block device elevator/IO-scheduler.
   3  *
   4  *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
   5  *
   6  * 30042000 Jens Axboe <axboe@kernel.dk> :
   7  *
   8  * Split the elevator a bit so that it is possible to choose a different
   9  * one or even write a new "plug in". There are three pieces:
  10  * - elevator_fn, inserts a new request in the queue list
  11  * - elevator_merge_fn, decides whether a new buffer can be merged with
  12  *   an existing request
  13  * - elevator_dequeue_fn, called when a request is taken off the active list
  14  *
  15  * 20082000 Dave Jones <davej@suse.de> :
  16  * Removed tests for max-bomb-segments, which was breaking elvtune
  17  *  when run without -bN
  18  *
  19  * Jens:
  20  * - Rework again to work with bio instead of buffer_heads
  21  * - loose bi_dev comparisons, partition handling is right now
  22  * - completely modularize elevator setup and teardown
  23  *
  24  */
  25 #include <linux/kernel.h>
  26 #include <linux/fs.h>
  27 #include <linux/blkdev.h>
  28 #include <linux/elevator.h>
  29 #include <linux/bio.h>
  30 #include <linux/module.h>
  31 #include <linux/slab.h>
  32 #include <linux/init.h>
  33 #include <linux/compiler.h>
  34 #include <linux/blktrace_api.h>
  35 #include <linux/hash.h>
  36 #include <linux/uaccess.h>
  37 #include <linux/pm_runtime.h>
  38 #include <linux/blk-cgroup.h>
  39
  40 #include <trace/events/block.h>
  41
  42 #include "blk.h"
  43 #include "blk-mq-sched.h"
  44
  45 static DEFINE_SPINLOCK(elv_list_lock);
  46 static LIST_HEAD(elv_list);
  47
  48 /*
  49  * Merge hash stuff.
  50  */
  51 #define rq_hash_key(rq)         (blk_rq_pos(rq) + blk_rq_sectors(rq))
  52
  53 /*
  54  * Query io scheduler to see if the current process issuing bio may be
  55  * merged with rq.
  56  */
  57 static int elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio)
  58 {
  59         struct request_queue *q = rq->q;
  60         struct elevator_queue *e = q->elevator;
  61
  62         if (e->uses_mq && e->type->ops.mq.allow_merge)
  63                 return e->type->ops.mq.allow_merge(q, rq, bio);
  64         else if (!e->uses_mq && e->type->ops.sq.elevator_allow_bio_merge_fn)
  65                 return e->type->ops.sq.elevator_allow_bio_merge_fn(q, rq, bio);
  66
  67         return 1;
  68 }
  69
  70 /*
  71  * can we safely merge with this request?
  72  */
  73 bool elv_bio_merge_ok(struct request *rq, struct bio *bio)
  74 {
  75         if (!blk_rq_merge_ok(rq, bio))
  76                 return false;
  77
  78         if (!elv_iosched_allow_bio_merge(rq, bio))
  79                 return false;
  80
  81         return true;
  82 }
  83 EXPORT_SYMBOL(elv_bio_merge_ok);
  84
  85 static struct elevator_type *elevator_find(const char *name)
  86 {
  87         struct elevator_type *e;
  88
  89         list_for_each_entry(e, &elv_list, list) {
  90                 if (!strcmp(e->elevator_name, name))
  91                         return e;
  92         }
  93
  94         return NULL;
  95 }
  96
  97 static void elevator_put(struct elevator_type *e)
  98 {
  99         module_put(e->elevator_owner);
 100 }
 101
 102 static struct elevator_type *elevator_get(const char *name, bool try_loading)
 103 {
 104         struct elevator_type *e;
 105
 106         spin_lock(&elv_list_lock);
 107
 108         e = elevator_find(name);
 109         if (!e && try_loading) {
 110                 spin_unlock(&elv_list_lock);
 111                 request_module("%s-iosched", name);
 112                 spin_lock(&elv_list_lock);
 113                 e = elevator_find(name);
 114         }
 115
 116         if (e && !try_module_get(e->elevator_owner))
 117                 e = NULL;
 118
 119         spin_unlock(&elv_list_lock);
 120
 121         return e;
 122 }
 123
 124 static char chosen_elevator[ELV_NAME_MAX];
 125
 126 static int __init elevator_setup(char *str)
 127 {
 128         /*
 129          * Be backwards-compatible with previous kernels, so users
 130          * won't get the wrong elevator.
 131          */
 132         strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 133         return 1;
 134 }
 135
 136 __setup("elevator=", elevator_setup);
 137
 138 /* called during boot to load the elevator chosen by the elevator param */
 139 void __init load_default_elevator_module(void)
 140 {
 141         struct elevator_type *e;
 142
 143         if (!chosen_elevator[0])
 144                 return;
 145
 146         spin_lock(&elv_list_lock);
 147         e = elevator_find(chosen_elevator);
 148         spin_unlock(&elv_list_lock);
 149
 150         if (!e)
 151                 request_module("%s-iosched", chosen_elevator);
 152 }
 153
 154 static struct kobj_type elv_ktype;
 155
 156 struct elevator_queue *elevator_alloc(struct request_queue *q,
 157                                   struct elevator_type *e)
 158 {
 159         struct elevator_queue *eq;
 160
 161         eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
 162         if (unlikely(!eq))
 163                 return NULL;
 164
 165         eq->type = e;
 166         kobject_init(&eq->kobj, &elv_ktype);
 167         mutex_init(&eq->sysfs_lock);
 168         hash_init(eq->hash);
 169         eq->uses_mq = e->uses_mq;
 170
 171         return eq;
 172 }
 173 EXPORT_SYMBOL(elevator_alloc);
 174
 175 static void elevator_release(struct kobject *kobj)
 176 {
 177         struct elevator_queue *e;
 178
 179         e = container_of(kobj, struct elevator_queue, kobj);
 180         elevator_put(e->type);
 181         kfree(e);
 182 }
 183
 184 int elevator_init(struct request_queue *q, char *name)
 185 {
 186         struct elevator_type *e = NULL;
 187         int err;
 188
 189         /*
 190          * q->sysfs_lock must be held to provide mutual exclusion between
 191          * elevator_switch() and here.
 192          */
 193         lockdep_assert_held(&q->sysfs_lock);
 194
 195         if (unlikely(q->elevator))
 196                 return 0;
 197
 198         INIT_LIST_HEAD(&q->queue_head);
 199         q->last_merge = NULL;
 200         q->end_sector = 0;
 201         q->boundary_rq = NULL;
 202
 203         if (name) {
 204                 e = elevator_get(name, true);
 205                 if (!e)
 206                         return -EINVAL;
 207         }
 208
 209         /*
 210          * Use the default elevator specified by config boot param or
 211          * config option.  Don't try to load modules as we could be running
 212          * off async and request_module() isn't allowed from async.
 213          */
 214         if (!e && *chosen_elevator) {
 215                 e = elevator_get(chosen_elevator, false);
 216                 if (!e)
 217                         printk(KERN_ERR "I/O scheduler %s not found\n",
 218                                                         chosen_elevator);
 219         }
 220
 221         if (!e) {
 222                 e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
 223                 if (!e) {
 224                         printk(KERN_ERR
 225                                 "Default I/O scheduler not found. " \
 226                                 "Using noop/none.\n");
 227                         if (q->mq_ops) {
 228                                 elevator_put(e);
 229                                 return 0;
 230                         }
 231                         e = elevator_get("noop", false);
 232                 }
 233         }
 234
 235         if (e->uses_mq) {
 236                 err = blk_mq_sched_setup(q);
 237                 if (!err)
 238                         err = e->ops.mq.init_sched(q, e);
 239         } else
 240                 err = e->ops.sq.elevator_init_fn(q, e);
 241         if (err) {
 242                 if (e->uses_mq)
 243                         blk_mq_sched_teardown(q);
 244                 elevator_put(e);
 245         }
 246         return err;
 247 }
 248 EXPORT_SYMBOL(elevator_init);
 249
 250 void elevator_exit(struct elevator_queue *e)
 251 {
 252         mutex_lock(&e->sysfs_lock);
 253         if (e->uses_mq && e->type->ops.mq.exit_sched)
 254                 e->type->ops.mq.exit_sched(e);
 255         else if (!e->uses_mq && e->type->ops.sq.elevator_exit_fn)
 256                 e->type->ops.sq.elevator_exit_fn(e);
 257         mutex_unlock(&e->sysfs_lock);
 258
 259         kobject_put(&e->kobj);
 260 }
 261 EXPORT_SYMBOL(elevator_exit);
 262
 263 static inline void __elv_rqhash_del(struct request *rq)
 264 {
 265         hash_del(&rq->hash);
 266         rq->rq_flags &= ~RQF_HASHED;
 267 }
 268
 269 void elv_rqhash_del(struct request_queue *q, struct request *rq)
 270 {
 271         if (ELV_ON_HASH(rq))
 272                 __elv_rqhash_del(rq);
 273 }
 274 EXPORT_SYMBOL_GPL(elv_rqhash_del);
 275
 276 void elv_rqhash_add(struct request_queue *q, struct request *rq)
 277 {
 278         struct elevator_queue *e = q->elevator;
 279
 280         BUG_ON(ELV_ON_HASH(rq));
 281         hash_add(e->hash, &rq->hash, rq_hash_key(rq));
 282         rq->rq_flags |= RQF_HASHED;
 283 }
 284 EXPORT_SYMBOL_GPL(elv_rqhash_add);
 285
 286 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 287 {
 288         __elv_rqhash_del(rq);
 289         elv_rqhash_add(q, rq);
 290 }
 291
 292 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 293 {
 294         struct elevator_queue *e = q->elevator;
 295         struct hlist_node *next;
 296         struct request *rq;
 297
 298         hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
 299                 BUG_ON(!ELV_ON_HASH(rq));
 300
 301                 if (unlikely(!rq_mergeable(rq))) {
 302                         __elv_rqhash_del(rq);
 303                         continue;
 304                 }
 305
 306                 if (rq_hash_key(rq) == offset)
 307                         return rq;
 308         }
 309
 310         return NULL;
 311 }
 312
 313 /*
 314  * RB-tree support functions for inserting/lookup/removal of requests
 315  * in a sorted RB tree.
 316  */
 317 void elv_rb_add(struct rb_root *root, struct request *rq)
 318 {
 319         struct rb_node **p = &root->rb_node;
 320         struct rb_node *parent = NULL;
 321         struct request *__rq;
 322
 323         while (*p) {
 324                 parent = *p;
 325                 __rq = rb_entry(parent, struct request, rb_node);
 326
 327                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 328                         p = &(*p)->rb_left;
 329                 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
 330                         p = &(*p)->rb_right;
 331         }
 332
 333         rb_link_node(&rq->rb_node, parent, p);
 334         rb_insert_color(&rq->rb_node, root);
 335 }
 336 EXPORT_SYMBOL(elv_rb_add);
 337
 338 void elv_rb_del(struct rb_root *root, struct request *rq)
 339 {
 340         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 341         rb_erase(&rq->rb_node, root);
 342         RB_CLEAR_NODE(&rq->rb_node);
 343 }
 344 EXPORT_SYMBOL(elv_rb_del);
 345
 346 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 347 {
 348         struct rb_node *n = root->rb_node;
 349         struct request *rq;
 350
 351         while (n) {
 352                 rq = rb_entry(n, struct request, rb_node);
 353
 354                 if (sector < blk_rq_pos(rq))
 355                         n = n->rb_left;
 356                 else if (sector > blk_rq_pos(rq))
 357                         n = n->rb_right;
 358                 else
 359                         return rq;
 360         }
 361
 362         return NULL;
 363 }
 364 EXPORT_SYMBOL(elv_rb_find);
 365
 366 /*
 367  * Insert rq into dispatch queue of q.  Queue lock must be held on
 368  * entry.  rq is sort instead into the dispatch queue. To be used by
 369  * specific elevators.
 370  */
 371 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 372 {
 373         sector_t boundary;
 374         struct list_head *entry;
 375
 376         if (q->last_merge == rq)
 377                 q->last_merge = NULL;
 378
 379         elv_rqhash_del(q, rq);
 380
 381         q->nr_sorted--;
 382
 383         boundary = q->end_sector;
 384         list_for_each_prev(entry, &q->queue_head) {
 385                 struct request *pos = list_entry_rq(entry);
 386
 387                 if (req_op(rq) != req_op(pos))
 388                         break;
 389                 if (rq_data_dir(rq) != rq_data_dir(pos))
 390                         break;
 391                 if (pos->rq_flags & (RQF_STARTED | RQF_SOFTBARRIER))
 392                         break;
 393                 if (blk_rq_pos(rq) >= boundary) {
 394                         if (blk_rq_pos(pos) < boundary)
 395                                 continue;
 396                 } else {
 397                         if (blk_rq_pos(pos) >= boundary)
 398                                 break;
 399                 }
 400                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 401                         break;
 402         }
 403
 404         list_add(&rq->queuelist, entry);
 405 }
 406 EXPORT_SYMBOL(elv_dispatch_sort);
 407
 408 /*
 409  * Insert rq into dispatch queue of q.  Queue lock must be held on
 410  * entry.  rq is added to the back of the dispatch queue. To be used by
 411  * specific elevators.
 412  */
 413 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 414 {
 415         if (q->last_merge == rq)
 416                 q->last_merge = NULL;
 417
 418         elv_rqhash_del(q, rq);
 419
 420         q->nr_sorted--;
 421
 422         q->end_sector = rq_end_sector(rq);
 423         q->boundary_rq = rq;
 424         list_add_tail(&rq->queuelist, &q->queue_head);
 425 }
 426 EXPORT_SYMBOL(elv_dispatch_add_tail);
 427
 428 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 429 {
 430         struct elevator_queue *e = q->elevator;
 431         struct request *__rq;
 432         int ret;
 433
 434         /*
 435          * Levels of merges:
 436          *      nomerges:  No merges at all attempted
 437          *      noxmerges: Only simple one-hit cache try
 438          *      merges:    All merge tries attempted
 439          */
 440         if (blk_queue_nomerges(q) || !bio_mergeable(bio))
 441                 return ELEVATOR_NO_MERGE;
 442
 443         /*
 444          * First try one-hit cache.
 445          */
 446         if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
 447                 ret = blk_try_merge(q->last_merge, bio);
 448                 if (ret != ELEVATOR_NO_MERGE) {
 449                         *req = q->last_merge;
 450                         return ret;
 451                 }
 452         }
 453
 454         if (blk_queue_noxmerges(q))
 455                 return ELEVATOR_NO_MERGE;
 456
 457         /*
 458          * See if our hash lookup can find a potential backmerge.
 459          */
 460         __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
 461         if (__rq && elv_bio_merge_ok(__rq, bio)) {
 462                 *req = __rq;
 463                 return ELEVATOR_BACK_MERGE;
 464         }
 465
 466         if (e->uses_mq && e->type->ops.mq.request_merge)
 467                 return e->type->ops.mq.request_merge(q, req, bio);
 468         else if (!e->uses_mq && e->type->ops.sq.elevator_merge_fn)
 469                 return e->type->ops.sq.elevator_merge_fn(q, req, bio);
 470
 471         return ELEVATOR_NO_MERGE;
 472 }
 473
 474 /*
 475  * Attempt to do an insertion back merge. Only check for the case where
 476  * we can append 'rq' to an existing request, so we can throw 'rq' away
 477  * afterwards.
 478  *
 479  * Returns true if we merged, false otherwise
 480  */
 481 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
 482 {
 483         struct request *__rq;
 484         bool ret;
 485
 486         if (blk_queue_nomerges(q))
 487                 return false;
 488
 489         /*
 490          * First try one-hit cache.
 491          */
 492         if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
 493                 return true;
 494
 495         if (blk_queue_noxmerges(q))
 496                 return false;
 497
 498         ret = false;
 499         /*
 500          * See if our hash lookup can find a potential backmerge.
 501          */
 502         while (1) {
 503                 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
 504                 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
 505                         break;
 506
 507                 /* The merged request could be merged with others, try again */
 508                 ret = true;
 509                 rq = __rq;
 510         }
 511
 512         return ret;
 513 }
 514
 515 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 516 {
 517         struct elevator_queue *e = q->elevator;
 518
 519         if (e->uses_mq && e->type->ops.mq.request_merged)
 520                 e->type->ops.mq.request_merged(q, rq, type);
 521         else if (!e->uses_mq && e->type->ops.sq.elevator_merged_fn)
 522                 e->type->ops.sq.elevator_merged_fn(q, rq, type);
 523
 524         if (type == ELEVATOR_BACK_MERGE)
 525                 elv_rqhash_reposition(q, rq);
 526
 527         q->last_merge = rq;
 528 }
 529
 530 void elv_merge_requests(struct request_queue *q, struct request *rq,
 531                              struct request *next)
 532 {
 533         struct elevator_queue *e = q->elevator;
 534         bool next_sorted = false;
 535
 536         if (e->uses_mq && e->type->ops.mq.requests_merged)
 537                 e->type->ops.mq.requests_merged(q, rq, next);
 538         else if (e->type->ops.sq.elevator_merge_req_fn) {
 539                 next_sorted = next->rq_flags & RQF_SORTED;
 540                 if (next_sorted)
 541                         e->type->ops.sq.elevator_merge_req_fn(q, rq, next);
 542         }
 543
 544         elv_rqhash_reposition(q, rq);
 545
 546         if (next_sorted) {
 547                 elv_rqhash_del(q, next);
 548                 q->nr_sorted--;
 549         }
 550
 551         q->last_merge = rq;
 552 }
 553
 554 void elv_bio_merged(struct request_queue *q, struct request *rq,
 555                         struct bio *bio)
 556 {
 557         struct elevator_queue *e = q->elevator;
 558
 559         if (WARN_ON_ONCE(e->uses_mq))
 560                 return;
 561
 562         if (e->type->ops.sq.elevator_bio_merged_fn)
 563                 e->type->ops.sq.elevator_bio_merged_fn(q, rq, bio);
 564 }
 565
 566 #ifdef CONFIG_PM
 567 static void blk_pm_requeue_request(struct request *rq)
 568 {
 569         if (rq->q->dev && !(rq->rq_flags & RQF_PM))
 570                 rq->q->nr_pending--;
 571 }
 572
 573 static void blk_pm_add_request(struct request_queue *q, struct request *rq)
 574 {
 575         if (q->dev && !(rq->rq_flags & RQF_PM) && q->nr_pending++ == 0 &&
 576             (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
 577                 pm_request_resume(q->dev);
 578 }
 579 #else
 580 static inline void blk_pm_requeue_request(struct request *rq) {}
 581 static inline void blk_pm_add_request(struct request_queue *q,
 582                                       struct request *rq)
 583 {
 584 }
 585 #endif
 586
 587 void elv_requeue_request(struct request_queue *q, struct request *rq)
 588 {
 589         /*
 590          * it already went through dequeue, we need to decrement the
 591          * in_flight count again
 592          */
 593         if (blk_account_rq(rq)) {
 594                 q->in_flight[rq_is_sync(rq)]--;
 595                 if (rq->rq_flags & RQF_SORTED)
 596                         elv_deactivate_rq(q, rq);
 597         }
 598
 599         rq->rq_flags &= ~RQF_STARTED;
 600
 601         blk_pm_requeue_request(rq);
 602
 603         __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
 604 }
 605
 606 void elv_drain_elevator(struct request_queue *q)
 607 {
 608         struct elevator_queue *e = q->elevator;
 609         static int printed;
 610
 611         if (WARN_ON_ONCE(e->uses_mq))
 612                 return;
 613
 614         lockdep_assert_held(q->queue_lock);
 615
 616         while (e->type->ops.sq.elevator_dispatch_fn(q, 1))
 617                 ;
 618         if (q->nr_sorted && printed++ < 10) {
 619                 printk(KERN_ERR "%s: forced dispatching is broken "
 620                        "(nr_sorted=%u), please report this\n",
 621                        q->elevator->type->elevator_name, q->nr_sorted);
 622         }
 623 }
 624
 625 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
 626 {
 627         trace_block_rq_insert(q, rq);
 628
 629         blk_pm_add_request(q, rq);
 630
 631         rq->q = q;
 632
 633         if (rq->rq_flags & RQF_SOFTBARRIER) {
 634                 /* barriers are scheduling boundary, update end_sector */
 635                 if (rq->cmd_type == REQ_TYPE_FS) {
 636                         q->end_sector = rq_end_sector(rq);
 637                         q->boundary_rq = rq;
 638                 }
 639         } else if (!(rq->rq_flags & RQF_ELVPRIV) &&
 640                     (where == ELEVATOR_INSERT_SORT ||
 641                      where == ELEVATOR_INSERT_SORT_MERGE))
 642                 where = ELEVATOR_INSERT_BACK;
 643
 644         switch (where) {
 645         case ELEVATOR_INSERT_REQUEUE:
 646         case ELEVATOR_INSERT_FRONT:
 647                 rq->rq_flags |= RQF_SOFTBARRIER;
 648                 list_add(&rq->queuelist, &q->queue_head);
 649                 break;
 650
 651         case ELEVATOR_INSERT_BACK:
 652                 rq->rq_flags |= RQF_SOFTBARRIER;
 653                 elv_drain_elevator(q);
 654                 list_add_tail(&rq->queuelist, &q->queue_head);
 655                 /*
 656                  * We kick the queue here for the following reasons.
 657                  * - The elevator might have returned NULL previously
 658                  *   to delay requests and returned them now.  As the
 659                  *   queue wasn't empty before this request, ll_rw_blk
 660                  *   won't run the queue on return, resulting in hang.
 661                  * - Usually, back inserted requests won't be merged
 662                  *   with anything.  There's no point in delaying queue
 663                  *   processing.
 664                  */
 665                 __blk_run_queue(q);
 666                 break;
 667
 668         case ELEVATOR_INSERT_SORT_MERGE:
 669                 /*
 670                  * If we succeed in merging this request with one in the
 671                  * queue already, we are done - rq has now been freed,
 672                  * so no need to do anything further.
 673                  */
 674                 if (elv_attempt_insert_merge(q, rq))
 675                         break;
 676         case ELEVATOR_INSERT_SORT:
 677                 BUG_ON(rq->cmd_type != REQ_TYPE_FS);
 678                 rq->rq_flags |= RQF_SORTED;
 679                 q->nr_sorted++;
 680                 if (rq_mergeable(rq)) {
 681                         elv_rqhash_add(q, rq);
 682                         if (!q->last_merge)
 683                                 q->last_merge = rq;
 684                 }
 685
 686                 /*
 687                  * Some ioscheds (cfq) run q->request_fn directly, so
 688                  * rq cannot be accessed after calling
 689                  * elevator_add_req_fn.
 690                  */
 691                 q->elevator->type->ops.sq.elevator_add_req_fn(q, rq);
 692                 break;
 693
 694         case ELEVATOR_INSERT_FLUSH:
 695                 rq->rq_flags |= RQF_SOFTBARRIER;
 696                 blk_insert_flush(rq);
 697                 break;
 698         default:
 699                 printk(KERN_ERR "%s: bad insertion point %d\n",
 700                        __func__, where);
 701                 BUG();
 702         }
 703 }
 704 EXPORT_SYMBOL(__elv_add_request);
 705
 706 void elv_add_request(struct request_queue *q, struct request *rq, int where)
 707 {
 708         unsigned long flags;
 709
 710         spin_lock_irqsave(q->queue_lock, flags);
 711         __elv_add_request(q, rq, where);
 712         spin_unlock_irqrestore(q->queue_lock, flags);
 713 }
 714 EXPORT_SYMBOL(elv_add_request);
 715
 716 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 717 {
 718         struct elevator_queue *e = q->elevator;
 719
 720         if (e->uses_mq && e->type->ops.mq.next_request)
 721                 return e->type->ops.mq.next_request(q, rq);
 722         else if (!e->uses_mq && e->type->ops.sq.elevator_latter_req_fn)
 723                 return e->type->ops.sq.elevator_latter_req_fn(q, rq);
 724
 725         return NULL;
 726 }
 727
 728 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 729 {
 730         struct elevator_queue *e = q->elevator;
 731
 732         if (e->uses_mq && e->type->ops.mq.former_request)
 733                 return e->type->ops.mq.former_request(q, rq);
 734         if (!e->uses_mq && e->type->ops.sq.elevator_former_req_fn)
 735                 return e->type->ops.sq.elevator_former_req_fn(q, rq);
 736         return NULL;
 737 }
 738
 739 int elv_set_request(struct request_queue *q, struct request *rq,
 740                     struct bio *bio, gfp_t gfp_mask)
 741 {
 742         struct elevator_queue *e = q->elevator;
 743
 744         if (WARN_ON_ONCE(e->uses_mq))
 745                 return 0;
 746
 747         if (e->type->ops.sq.elevator_set_req_fn)
 748                 return e->type->ops.sq.elevator_set_req_fn(q, rq, bio, gfp_mask);
 749         return 0;
 750 }
 751
 752 void elv_put_request(struct request_queue *q, struct request *rq)
 753 {
 754         struct elevator_queue *e = q->elevator;
 755
 756         if (WARN_ON_ONCE(e->uses_mq))
 757                 return;
 758
 759         if (e->type->ops.sq.elevator_put_req_fn)
 760                 e->type->ops.sq.elevator_put_req_fn(rq);
 761 }
 762
 763 int elv_may_queue(struct request_queue *q, unsigned int op)
 764 {
 765         struct elevator_queue *e = q->elevator;
 766
 767         if (WARN_ON_ONCE(e->uses_mq))
 768                 return 0;
 769
 770         if (e->type->ops.sq.elevator_may_queue_fn)
 771                 return e->type->ops.sq.elevator_may_queue_fn(q, op);
 772
 773         return ELV_MQUEUE_MAY;
 774 }
 775
 776 void elv_completed_request(struct request_queue *q, struct request *rq)
 777 {
 778         struct elevator_queue *e = q->elevator;
 779
 780         if (WARN_ON_ONCE(e->uses_mq))
 781                 return;
 782
 783         /*
 784          * request is released from the driver, io must be done
 785          */
 786         if (blk_account_rq(rq)) {
 787                 q->in_flight[rq_is_sync(rq)]--;
 788                 if ((rq->rq_flags & RQF_SORTED) &&
 789                     e->type->ops.sq.elevator_completed_req_fn)
 790                         e->type->ops.sq.elevator_completed_req_fn(q, rq);
 791         }
 792 }
 793
 794 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 795
 796 static ssize_t
 797 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 798 {
 799         struct elv_fs_entry *entry = to_elv(attr);
 800         struct elevator_queue *e;
 801         ssize_t error;
 802
 803         if (!entry->show)
 804                 return -EIO;
 805
 806         e = container_of(kobj, struct elevator_queue, kobj);
 807         mutex_lock(&e->sysfs_lock);
 808         error = e->type ? entry->show(e, page) : -ENOENT;
 809         mutex_unlock(&e->sysfs_lock);
 810         return error;
 811 }
 812
 813 static ssize_t
 814 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 815                const char *page, size_t length)
 816 {
 817         struct elv_fs_entry *entry = to_elv(attr);
 818         struct elevator_queue *e;
 819         ssize_t error;
 820
 821         if (!entry->store)
 822                 return -EIO;
 823
 824         e = container_of(kobj, struct elevator_queue, kobj);
 825         mutex_lock(&e->sysfs_lock);
 826         error = e->type ? entry->store(e, page, length) : -ENOENT;
 827         mutex_unlock(&e->sysfs_lock);
 828         return error;
 829 }
 830
 831 static const struct sysfs_ops elv_sysfs_ops = {
 832         .show   = elv_attr_show,
 833         .store  = elv_attr_store,
 834 };
 835
 836 static struct kobj_type elv_ktype = {
 837         .sysfs_ops      = &elv_sysfs_ops,
 838         .release        = elevator_release,
 839 };
 840
 841 int elv_register_queue(struct request_queue *q)
 842 {
 843         struct elevator_queue *e = q->elevator;
 844         int error;
 845
 846         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 847         if (!error) {
 848                 struct elv_fs_entry *attr = e->type->elevator_attrs;
 849                 if (attr) {
 850                         while (attr->attr.name) {
 851                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 852                                         break;
 853                                 attr++;
 854                         }
 855                 }
 856                 kobject_uevent(&e->kobj, KOBJ_ADD);
 857                 e->registered = 1;
 858                 if (!e->uses_mq && e->type->ops.sq.elevator_registered_fn)
 859                         e->type->ops.sq.elevator_registered_fn(q);
 860         }
 861         return error;
 862 }
 863 EXPORT_SYMBOL(elv_register_queue);
 864
 865 void elv_unregister_queue(struct request_queue *q)
 866 {
 867         if (q) {
 868                 struct elevator_queue *e = q->elevator;
 869
 870                 kobject_uevent(&e->kobj, KOBJ_REMOVE);
 871                 kobject_del(&e->kobj);
 872                 e->registered = 0;
 873         }
 874 }
 875 EXPORT_SYMBOL(elv_unregister_queue);
 876
 877 int elv_register(struct elevator_type *e)
 878 {
 879         char *def = "";
 880
 881         /* create icq_cache if requested */
 882         if (e->icq_size) {
 883                 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
 884                     WARN_ON(e->icq_align < __alignof__(struct io_cq)))
 885                         return -EINVAL;
 886
 887                 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
 888                          "%s_io_cq", e->elevator_name);
 889                 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
 890                                                  e->icq_align, 0, NULL);
 891                 if (!e->icq_cache)
 892                         return -ENOMEM;
 893         }
 894
 895         /* register, don't allow duplicate names */
 896         spin_lock(&elv_list_lock);
 897         if (elevator_find(e->elevator_name)) {
 898                 spin_unlock(&elv_list_lock);
 899                 if (e->icq_cache)
 900                         kmem_cache_destroy(e->icq_cache);
 901                 return -EBUSY;
 902         }
 903         list_add_tail(&e->list, &elv_list);
 904         spin_unlock(&elv_list_lock);
 905
 906         /* print pretty message */
 907         if (!strcmp(e->elevator_name, chosen_elevator) ||
 908                         (!*chosen_elevator &&
 909                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 910                                 def = " (default)";
 911
 912         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 913                                                                 def);
 914         return 0;
 915 }
 916 EXPORT_SYMBOL_GPL(elv_register);
 917
 918 void elv_unregister(struct elevator_type *e)
 919 {
 920         /* unregister */
 921         spin_lock(&elv_list_lock);
 922         list_del_init(&e->list);
 923         spin_unlock(&elv_list_lock);
 924
 925         /*
 926          * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
 927          * sure all RCU operations are complete before proceeding.
 928          */
 929         if (e->icq_cache) {
 930                 rcu_barrier();
 931                 kmem_cache_destroy(e->icq_cache);
 932                 e->icq_cache = NULL;
 933         }
 934 }
 935 EXPORT_SYMBOL_GPL(elv_unregister);
 936
 937 /*
 938  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 939  * we don't free the old io scheduler, before we have allocated what we
 940  * need for the new one. this way we have a chance of going back to the old
 941  * one, if the new one fails init for some reason.
 942  */
 943 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 944 {
 945         struct elevator_queue *old = q->elevator;
 946         bool old_registered = false;
 947         int err;
 948
 949         if (q->mq_ops) {
 950                 blk_mq_freeze_queue(q);
 951                 blk_mq_quiesce_queue(q);
 952         }
 953
 954         /*
 955          * Turn on BYPASS and drain all requests w/ elevator private data.
 956          * Block layer doesn't call into a quiesced elevator - all requests
 957          * are directly put on the dispatch list without elevator data
 958          * using INSERT_BACK.  All requests have SOFTBARRIER set and no
 959          * merge happens either.
 960          */
 961         if (old) {
 962                 old_registered = old->registered;
 963
 964                 if (old->uses_mq)
 965                         blk_mq_sched_teardown(q);
 966
 967                 if (!q->mq_ops)
 968                         blk_queue_bypass_start(q);
 969
 970                 /* unregister and clear all auxiliary data of the old elevator */
 971                 if (old_registered)
 972                         elv_unregister_queue(q);
 973
 974                 spin_lock_irq(q->queue_lock);
 975                 ioc_clear_queue(q);
 976                 spin_unlock_irq(q->queue_lock);
 977         }
 978
 979         /* allocate, init and register new elevator */
 980         if (new_e) {
 981                 if (new_e->uses_mq) {
 982                         err = blk_mq_sched_setup(q);
 983                         if (!err)
 984                                 err = new_e->ops.mq.init_sched(q, new_e);
 985                 } else
 986                         err = new_e->ops.sq.elevator_init_fn(q, new_e);
 987                 if (err)
 988                         goto fail_init;
 989
 990                 err = elv_register_queue(q);
 991                 if (err)
 992                         goto fail_register;
 993         } else
 994                 q->elevator = NULL;
 995
 996         /* done, kill the old one and finish */
 997         if (old) {
 998                 elevator_exit(old);
 999                 if (!q->mq_ops)
1000                         blk_queue_bypass_end(q);
1001         }
1002
1003         if (q->mq_ops) {
1004                 blk_mq_unfreeze_queue(q);
1005                 blk_mq_start_stopped_hw_queues(q, true);
1006         }
1007
1008         if (new_e)
1009                 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
1010         else
1011                 blk_add_trace_msg(q, "elv switch: none");
1012
1013         return 0;
1014
1015 fail_register:
1016         if (q->mq_ops)
1017                 blk_mq_sched_teardown(q);
1018         elevator_exit(q->elevator);
1019 fail_init:
1020         /* switch failed, restore and re-register old elevator */
1021         if (old) {
1022                 q->elevator = old;
1023                 elv_register_queue(q);
1024                 if (!q->mq_ops)
1025                         blk_queue_bypass_end(q);
1026         }
1027         if (q->mq_ops) {
1028                 blk_mq_unfreeze_queue(q);
1029                 blk_mq_start_stopped_hw_queues(q, true);
1030         }
1031
1032         return err;
1033 }
1034
1035 /*
1036  * Switch this queue to the given IO scheduler.
1037  */
1038 static int __elevator_change(struct request_queue *q, const char *name)
1039 {
1040         char elevator_name[ELV_NAME_MAX];
1041         struct elevator_type *e;
1042
1043         /*
1044          * Special case for mq, turn off scheduling
1045          */
1046         if (q->mq_ops && !strncmp(name, "none", 4))
1047                 return elevator_switch(q, NULL);
1048
1049         strlcpy(elevator_name, name, sizeof(elevator_name));
1050         e = elevator_get(strstrip(elevator_name), true);
1051         if (!e) {
1052                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1053                 return -EINVAL;
1054         }
1055
1056         if (q->elevator &&
1057             !strcmp(elevator_name, q->elevator->type->elevator_name)) {
1058                 elevator_put(e);
1059                 return 0;
1060         }
1061
1062         if (!e->uses_mq && q->mq_ops) {
1063                 elevator_put(e);
1064                 return -EINVAL;
1065         }
1066         if (e->uses_mq && !q->mq_ops) {
1067                 elevator_put(e);
1068                 return -EINVAL;
1069         }
1070
1071         return elevator_switch(q, e);
1072 }
1073
1074 int elevator_change(struct request_queue *q, const char *name)
1075 {
1076         int ret;
1077
1078         /* Protect q->elevator from elevator_init() */
1079         mutex_lock(&q->sysfs_lock);
1080         ret = __elevator_change(q, name);
1081         mutex_unlock(&q->sysfs_lock);
1082
1083         return ret;
1084 }
1085 EXPORT_SYMBOL(elevator_change);
1086
1087 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1088                           size_t count)
1089 {
1090         int ret;
1091
1092         if (!(q->mq_ops || q->request_fn))
1093                 return count;
1094
1095         ret = __elevator_change(q, name);
1096         if (!ret)
1097                 return count;
1098
1099         printk(KERN_ERR "elevator: switch to %s failed\n", name);
1100         return ret;
1101 }
1102
1103 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1104 {
1105         struct elevator_queue *e = q->elevator;
1106         struct elevator_type *elv = NULL;
1107         struct elevator_type *__e;
1108         int len = 0;
1109
1110         if (!blk_queue_stackable(q))
1111                 return sprintf(name, "none\n");
1112
1113         if (!q->elevator)
1114                 len += sprintf(name+len, "[none] ");
1115         else
1116                 elv = e->type;
1117
1118         spin_lock(&elv_list_lock);
1119         list_for_each_entry(__e, &elv_list, list) {
1120                 if (elv && !strcmp(elv->elevator_name, __e->elevator_name)) {
1121                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1122                         continue;
1123                 }
1124                 if (__e->uses_mq && q->mq_ops)
1125                         len += sprintf(name+len, "%s ", __e->elevator_name);
1126                 else if (!__e->uses_mq && !q->mq_ops)
1127                         len += sprintf(name+len, "%s ", __e->elevator_name);
1128         }
1129         spin_unlock(&elv_list_lock);
1130
1131         if (q->mq_ops && q->elevator)
1132                 len += sprintf(name+len, "none");
1133
1134         len += sprintf(len+name, "\n");
1135         return len;
1136 }
1137
1138 struct request *elv_rb_former_request(struct request_queue *q,
1139                                       struct request *rq)
1140 {
1141         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1142
1143         if (rbprev)
1144                 return rb_entry_rq(rbprev);
1145
1146         return NULL;
1147 }
1148 EXPORT_SYMBOL(elv_rb_former_request);
1149
1150 struct request *elv_rb_latter_request(struct request_queue *q,
1151                                       struct request *rq)
1152 {
1153         struct rb_node *rbnext = rb_next(&rq->rb_node);
1154
1155         if (rbnext)
1156                 return rb_entry_rq(rbnext);
1157
1158         return NULL;
1159 }
1160 EXPORT_SYMBOL(elv_rb_latter_request);